Improving Error Related Potential Classification by using Generative Adversarial Networks and Deep Convolutional Neural Networks

Abstract

An error-related potential (ErrP) is a form of event-related potential that occurs when an error-related stimulus is encountered during a task. The decoding of ErrP has the potential to apply for Brain Computer System (BCI). Although various methods have been applied to the decoding of ErrPs, existing classification methods have room for improvement. Using a deep convolutional neural network (DCNN) is a viable approach to ErrP classification but its performance can be compromised by insufficient training data being available. Using a form of generative adversarial network (GAN) enables data augmentation, which has offered significant performance improvement in a variety of fields such as signal processing, robotics, and unmanned vehicles. We therefore propose a novel approach to ErrP classification that combines GAN with DCNN to form a data-augmented DCNN. The proposed method has two main components: a GAN that offers data augmentation and a DCNN that performs the classification. We applied our method to the BNCI2020 dataset 22: Monitoring ErrPs, evaluating the results in terms of classification accuracy in various categories, including single-subject single-session, cross-subject single session, cross-session single-subject, and cross-subject cross-session versions for the entire dataset. The evaluations showed that the classification results had been improved comprehensively in comparison with existing published results. The maximum accuracy of the classification performance for the entire dataset reached 87%, which is 3% above the previous best result.